Energy-aware Real-Time Task Decomposition for partionned-EDF Scheduling on Multi-core Uniform Architectures

Modern multi-core embedded processors allow to implement increasingly complex processing applications, as any processing of large amount of data. Many of these are submitted to timing constraints, and can be implemented as parallel tasks by taking advantage of a multi-core architecture. However a careless implementation can waste precious energy, especially when processing elements are powered by batteries. Thus, It is necessary to calibrate the operational frequency and voltage of the processors so that the consumed energy is reduced while still meeting the timing requirements. In this paper, we explore the design choices in implementing parallel real-time applications on multi-core systems. We propose a realistic parallel real-time task model, NLP formulation and some heuristics for allocating threads to processors and selecting in offline their operational frequency. Experimental results with synthetic tasks sets show that it is time consuming to use NLP-solvers and that our heuristics are effective in reducing the power consumption